}

// -----

func.func @pack_CPU(%arg0: tensor<100xf32>, %arg1: tensor<100xf32>) -> tensor<2x100xf32> attributes {tac.device = "CPU", tac.interface_name = "func_2"} {
  // CHECK: tac.cost = 1.000000e+02
  %0 = "tfl.pack"(%arg0, %arg1) {axis = 0 : i32, tac.device = "CPU", values_count = 2 : i32} : (tensor<100xf32>, tensor<100xf32>) -> tensor<2x100xf32>
  func.return %0 : tensor<2x100xf32>
}

    %2 = func.call @func_2_CPU_FLOAT(%0, %1) {tac.device = "CPU", tac.inference_type = "FLOAT", tac.interface_name = "func_2"} : (tensor<100xf32>, tensor<100xf32>) -> tensor<2x100xf32>
    func.return %2 : tensor<2x100xf32>
  }

// CHECK-NEXT:  [[NEG_ONE_1:%.*]] = arith.constant dense<-1> : tensor<i32>
// CHECK-NEXT:  [[RESULT:%.*]] = "tf.Concat"([[ZERO]], [[INPUT]], [[EXTENDED_PART]]) : (tensor<i32>, tensor<3x10xf32>, tensor<?x10xf32>) -> tensor<?x10xf32>
// CHECK-NEXT:  return [[RESULT]] : tensor<?x10xf32>


// CHECK-LABEL:  func private @cond_false

    %18 = "tf.BiasAdd"(%17, %arg4) {data_format = "NHWC"} : (tensor<2x1001xf32>, tensor<1001xf32>) -> tensor<2x1001xf32>
    %19 = "tf.Reshape"(%18, %7) : (tensor<2x1001xf32>, tensor<2xi32>) -> tensor<2x1001xf32>
    %loss, %backprop = "tf.SparseSoftmaxCrossEntropyWithLogits"(%19, %13) : (tensor<2x1001xf32>, tensor<2xi64>) -> (tensor<2xf32>, tensor<2x1001xf32>)

func.func @optimizeTopKOnlyValues(%arg: tensor<3x10xf32>) -> tensor<3x5xf32>{
  %K = arith.constant dense<10> : tensor<i32>
  %values, %indices = "tfl.topk_v2"(%arg, %K) : (tensor<3x10xf32>, tensor<i32>) -> (tensor<3x10xf32>, tensor<3x10xi32>)
  %cst_0 = arith.constant dense<0> : tensor<2xi32>
  %cst_1 = arith.constant dense<[3, 5]> : tensor<2xi32>
  %0 = "tfl.slice"(%values, %cst_0, %cst_1) : (tensor<3x10xf32>, tensor<2xi32>, tensor<2xi32>) -> tensor<3x5xf32>

// CHECK:         }
func.func @broadcast_add(%arg0: tensor<1x1xf32>, %arg1: tensor<1x1000xf32>) -> (tensor<1x1000xf32>, tensor<1x1000xf32>) {
  %0 = "mhlo.broadcast_in_dim"(%arg0) <{broadcast_dimensions = dense<[0, 1]> : tensor<2xi64>}> : (tensor<1x1xf32>) -> tensor<1x1000xf32>
  %1 = mhlo.add %0, %arg1 : tensor<1x1000xf32>
  %2 = mhlo.add %arg1, %0 : tensor<1x1000xf32>

  %x = "tf.TensorArrayReadV3"(%handle_0, %index, %flow_0) {device = "/job:localhost/replica:0/task:0/device:CPU:0"} : (tensor<2x!tf_type.resource<tensor<?x100xf32>>>, tensor<i32>, tensor<f32>) -> tensor<?x100xf32>
  %y = "tf.MatMul"(%x, %cst) {device = "/job:localhost/replica:0/task:0/device:CPU:0"} : (tensor<?x100xf32>, tensor<100x512xf32>) -> (tensor<?x512xf32>)

  %x = "tf.TensorArrayReadV3"(%handle_0, %index, %flow_0) {device = "/job:localhost/replica:0/task:0/device:CPU:0"} : (tensor<2x!tf_type.resource<tensor<?x100xf32>>>, tensor<i32>, tensor<f32>) -> tensor<?x100xf32>
  %y = "tf.MatMul"(%x, %cst) {device = "/job:localhost/replica:0/task:0/device:CPU:0"} : (tensor<?x100xf32>, tensor<100x512xf32>) -> (tensor<?x512xf32>)

}

    %0 = "tf.MatMul"(%arg0, %arg1) {attr_map = "0:transpose_a,1:transpose_a", device = "", transpose_a = false, transpose_b = false} : (tensor<1x2x2x3xf32>, tensor<2x1024xf32>) -> tensor<*xf32>
    return %0 : tensor<*xf32>
  }

// CHECK: %[[cst:.*]] = "arith.constant"() <{value = dense<0.000000e+00> : tensor<2x1024xf32>}> : () -> tensor<2x1024xf32>